v2.0.0 shipped · MIT · works with 6 agents
One shared memory for
every AI coding agent.
Switch from Codex to Claude to Antigravity, or come back tomorrow after a usage limit, and your agent already knows what you're building, where you are, and what's next. No re-explaining.
irm https://raw.githubusercontent.com/AnasNafees1802/continuum/main/bootstrap.ps1 | iex
Installs once, for every agent on your machine. No runtime. No service. Plain Markdown + JSON.
Context is trapped in silos.
Every agent keeps its own private history, and none of them share.
😤 Without Continuum
You ship a couple of features, then the daily limit hits and your agent stops. You switch to another one, but it knows nothing. You burn time and tokens re-explaining the project, and you lose the thread of why things were done a certain way.
🔁 With Continuum
A small, structured memory ledger lives inside your repo at .aicontext/. Every agent follows one rule: read it when you start, update it before you stop. Because the memory isn't trapped in any one tool, context survives the switch.
Three moments, on autopilot.
Native hooks fire them for you, so continuity never depends on anyone remembering.
Catch up
A session-start hook injects STATE.md into context and verifies it against git. Your agent briefs you before you type a word.
Keep it live
Tasks and decisions stay current as you go. A pre-compaction hook prompts a handoff right as the context window fills, while the detail still exists.
Save the thread
One command stamps the ledger + git commit. And if a crash beats you to it, continuum import reconstructs the session from the transcript, plus git.
Open a fresh session. Ask one question.
No context pasted. No project explained. It just knows.
Built for the moment it matters.
The naïve version breaks exactly when a usage limit kills the session. Continuum doesn't rely on remembering.
Deterministic hooks
Catch-up, handoff, and a one-per-session nudge fire automatically, on Claude Code, Codex, Gemini and Cursor. Not a polite request to the model.
Reconstruction
A limit-killed session is rebuilt from the last agent's own transcript plus a git view, even across a Codex ⇄ Claude switch.
Never trusts a stale ledger
Catch-up compares the saved commit to HEAD and flags drift, so an agent verifies against git instead of briefing you wrong.
Universal git floor
Even where hooks or transcripts aren't available, git-based reconstruction rebuilds "what changed." Continuity never fully fails, on any tool.
Atomic, portable writes
Every ledger write is atomic and BOM-free, no truncated files, and JSON that every tool's parser can read. Cross-tool by construction.
Zero dependencies
Plain PowerShell + bash and Markdown + JSON. No runtime, no service, no lock-in. Readable and editable by hand.
What's automated, per agent.
Honest status, verified means confirmed on a real session.
| Agent | Catch-up | Handoff | Stop nudge | Reconstruction | Status |
|---|---|---|---|---|---|
| Claude Code | ✓ hook | ✓ | ✓ | native | ● verified |
| Codex | ✓ hook | ✓ | ✓ | native | ● verified |
| Antigravity | ✓ protocol | - | - | git | ● verified |
| Gemini CLI | ✓ hook | ✓ | - | native | ● beta |
| Cursor | ✓ hook | ✓ | ✓ | git | ● beta |
| Windsurf | ✓ per-turn | - | - | git | ● beta |
Every agent also participates via the honor-protocol (AGENTS.md / GEMINI.md / rules) and the universal git floor.
It saves tokens, not just time.
Real, structural savings. No fuzzy benchmarks, just less waste.
No re-explaining
Catch-up injects a compact summary (a few hundred tokens) instead of you re-pasting the whole project into a fresh session every time.
No redone work
A session killed by a usage limit gets reconstructed, not repeated. Redoing a whole session is the most expensive waste there is.
Bounded context
Journal rotation keeps the ledger small, so the memory never bloats the very context window it exists to save.
Up and running in one step.
Pick whichever suits you. All three do the same thing, and the memory ledger is created automatically the first time an agent opens a project.
Let your AI agent install it
EasiestInstall "Continuum" globally on my machine. It's an open-source (MIT) tool that gives AI coding agents a shared memory so they keep context across sessions and tools: https://github.com/AnasNafees1802/continuum 1. Detect my operating system. 2. On Windows, run in PowerShell: irm https://raw.githubusercontent.com/AnasNafees1802/continuum/main/bootstrap.ps1 | iex On macOS or Linux, run in bash: curl -fsSL https://raw.githubusercontent.com/AnasNafees1802/continuum/main/bootstrap.sh | bash 3. When it finishes, tell me which agents it set up and remind me to restart any open agent sessions. Do not change anything else.
Paste this into Claude Code, Codex, Cursor, Gemini or any coding agent. It will ask you to approve the command. Approve it, and you are done.
One line in your terminal
irm https://raw.githubusercontent.com/AnasNafees1802/continuum/main/bootstrap.ps1 | iex
Clone and run it yourself
git clone https://github.com/AnasNafees1802/continuum.git cd continuum # Windows: powershell -ExecutionPolicy Bypass -File .\install-global.ps1 # macOS/Linux: bash ./install-global.sh
Want to read the code before running anything? See all install options on GitHub.
Then open any project in any agent. It reads .aicontext/STATE.md and catches you up automatically.
Stop re-explaining your project.
Give every AI agent one shared brain. It's open source, MIT, and installs in one line.